RESEARCH ARTICLE

www.advancedscience.com

Green Algorithms: Quantifying the Carbon Footprint

of Computation
Loïc Lannelongue,* Jason Grealey, and Michael Inouye*

1. Introduction
Climate change is profoundly affecting nearly all aspects of life on earth,
including human societies, economies, and health. Various human activities The concentration of greenhouse gases
(GHGs) in the atmosphere has a dramatic
are responsible for significant greenhouse gas (GHG) emissions, including
influence on climate change with both
data centers and other sources of large-scale computation. Although many global and locally focused consequences,
important scientific milestones are achieved thanks to the development of such as rising sea levels, devastating wild-
high-performance computing, the resultant environmental impact is fires in Australia, extreme typhoons in the
underappreciated. In this work, a methodological framework to estimate the Pacific, severe droughts across Africa, as
carbon footprint of any computational task in a standardized and reliable way well as repercussions for human health.
With 100 megatonnes of CO2 emissions
is presented and metrics to contextualize GHG emissions are defined. A freely
per year (Note S1, Supporting information),
available online tool, Green Algorithms (www.green-algorithms.org) is similar to American commercial aviation,
developed, which enables a user to estimate and report the carbon footprint the contribution of data centers and high-
of their computation. The tool easily integrates with computational processes performance computing facilities to climate
as it requires minimal information and does not interfere with existing code, change is substantial. So far, rapidly in-
creasing demand has been paralleled by
while also accounting for a broad range of hardware configurations. Finally,
increasingly energy-efficient facilities, with
the GHG emissions of algorithms used for particle physics simulations, overall electricity consumption of data cen-
weather forecasts, and natural language processing are quantified. Taken ters somewhat stable. However, this stabil-
together, this study develops a simple generalizable framework and freely ity is likely to end in the coming years, with
available tool to quantify the carbon footprint of nearly any computation. a best-case scenario forecasting a three-
Combined with recommendations to minimize unnecessary CO2 emissions, fold increase in the energy needs of the
sector.[1,2]
the authors hope to raise awareness and facilitate greener computation.
Advances in computation, includ-
ing those in hardware, software, and

L. Lannelongue, Prof. M. Inouye J. Grealey, Prof. M. Inouye

Cambridge Baker Systems Genomics Initiative Cambridge Baker Systems Genomics Initiative
Department of Public Health and Primary Care Baker Heart and Diabetes Institute
University of Cambridge Melbourne, Victoria 3004, Australia
Cambridge CB1 8RN, UK J. Grealey
E-mail: LL582@medschl.cam.ac.uk; minouye@baker.edu.au; Department of Mathematics and Statistics
mi336@medschl.cam.ac.uk La Trobe University
L. Lannelongue, Prof. M. Inouye Melbourne 3086, Australia
British Heart Foundation Cardiovascular Epidemiology Unit Prof. M. Inouye
Department of Public Health and Primary Care British Heart Foundation Centre of Research Excellence
University of Cambridge University of Cambridge
Cambridge CB1 8RN, UK Cambridge CB1 8RN, UK
L. Lannelongue, Prof. M. Inouye Prof. M. Inouye
Health Data Research UK Cambridge National Institute for Health Research Cambridge Biomedical Research
Wellcome Genome Campus and University of Cambridge Centre
Cambridge CB10 1SA, UK University of Cambridge and Cambridge University Hospitals
Cambridge CB2 0QQ, UK
Prof. M. Inouye
The Alan Turing Institute
The ORCID identification number(s) for the author(s) of this article
London NW1 2DB, UK
can be found under https://doi.org/10.1002/advs.202100707
© 2021 The Authors. Advanced Science published by Wiley-VCH GmbH.
This is an open access article under the terms of the Creative Commons
Attribution License, which permits use, distribution and reproduction in
any medium, provided the original work is properly cited.
DOI: 10.1002/advs.202100707

Adv. Sci. 2021, 8, 2100707 2100707 (1 of 10) © 2021 The Authors. Advanced Science published by Wiley-VCH GmbH
www.advancedsciencenews.com
algorithms, have enabled scientific research to progress at un-
precedented rates. Weather forecasts have increased in accuracy
to the point where 5-day forecasts are approximately as accu-
rate as 1-day forecasts 40 years ago,[3] physics algorithms have
produced the first direct image of a black hole 55 million light-
years away,[4–6] the human genome has been mined to uncover
thousands of genetic variants for disease,[7] and machine learn-
ing (ML) permeates many aspects of society, including economic
and social interactions.[8–11] An example of the scale of compu-
tational research in science is the Extreme Science and Engi-
neering Discovery Environment (XSEDE) in the USA. In 2020
only, almost 9 billion compute hours have been used for sci-
entific computing,[12] a pace of 24 million h per day. Yet, the
costs associated with large-scale computation are not being fully
captured.
Power consumption results in GHG emissions and the envi-
ronmental costs of performing computations using data centers,
personal computers, and the immense diversity of architectures
are unclear. While programmes in green computing (the study of
environmentally responsible information and communications
technologies) have been developed over the past decade, these
mainly focus on energy-efficient hardware and cloud-related
technologies.[13–15]
With widely recognized power-hungry and expensive train-
ing algorithms, deep learning has begun to address its carbon
footprint. ML models have grown exponentially in size over the
past few years,[16] with some algorithms training for thousands
of core-hours, and the associated energy consumption and cost
have become a growing concern.[17] In natural language process-
ing (NLP), Strubell et al.[18] found that designing and training
translation engines can emit between 0.6 and 280 tonnes of CO2 .
While not all NLP algorithms require frequent retraining, algo-
rithms in other fields are run daily or weekly, multiplying their
energy consumption. Astronomy also relies largely on supercom-
puters to analyses data, which has motivated some investigations
into the carbon footprint of the field.[19,20] For example, it has
been estimated that the usage of supercomputers by Australian
astronomers emits 15 kilotonnes of CO2 per year, equivalent to
22 tonnes per researcher.[21] Cryptocurrencies, and their so-called
“mining farms,” have also seen their environmental impact in-
crease exponentially in recent years, and several reports have
shed doubts on their sustainability. A 2018 study estimated the
yearly energy usage of Bitcoin to be 46 TWh, resulting in 22 Mt of
CO2 released in the atmosphere.[22] In March 2021, Bitcoin usage
is estimated to be 130 TWh, which, if Bitcoin was a country, would
rank its energy usage in 28th highest position in the world, ahead
of Argentina and Ukraine.[23] Although crypto-mining relies on
dedicated hardware (application-specific integrated circuits) in-
stead of usual processors; therefore, it does not compete directly
with scientific computing. Regardless, the magnitude of its car-
bon footprint needs to be addressed urgently.
Previous studies have made advances in estimating GHG
emissions of computation but have limitations which preclude
broad applicability. These limitations include the requirement
that users self-monitor their power consumption[18] and are
restricted with respect to hardware (e.g., GPUs and/or cloud
systems[24,25] ), software (e.g., Python package integration[25] ), or
applications (e.g., ML).[18,24,25] To facilitate green computing and
widespread user uptake, there is a clear, and arguably urgent,
Adv. Sci. 2021, 8, 2100707 2100707 (2 of 10)
www.advancedscience.com
need for both a general and easy-to-use methodology for estimat-
ing carbon emissions that can be applied to any computational
task.
In this study, we present a simple and widely applicable
method and a tool for estimating the carbon footprint of com-
putation. The method considers the different sources of energy
usage, such as processors and memory, overhead of comput-
ing facilities and geographic location, while balancing accuracy
and practicality. The online calculator (www.green-algorithms.
org) implements this methodology and provides further context
by interpreting carbon amounts using travel distances and car-
bon sequestration. We demonstrate the applicability of the Green
Algorithms method by estimating the carbon footprint of parti-
cle physics simulations, weather forecast models, and NLP al-
gorithms as well as the carbon effects of distributed computa-
tion using multiple CPUs. Finally, we make recommendations
on ways for scientists to reduce their GHG emissions as well as
discuss the limitations of our approach.
2. Results
We developed a simple method which estimates the carbon foot-
print of an algorithm based on a number of factors, including
the hardware requirements of the tool, the runtime and the loca-
tion of the data center (Experimental Section). Using a pragmatic
scaling factor (PSF), we further augment our model by allowing
for empirical estimates of repeated computations for a particular
task, for example, parameter tuning and trial-and-errors. The re-
sultant gCO2 e is compared to the amount of carbon sequestered
by trees and the emissions of common activities such as driving
a car and air travel. We designed a freely available online tool,
Green Algorithms (www.green-algorithms.org; Figure 1), which
implements our approach and allows users to evaluate their com-
putations or estimate the carbon savings or costs of redeploying
them on other architectures.
We apply this tool to a range of algorithms selected from a va-
riety of scientific fields: physics (particle simulations and DNA
irradiation), atmospheric sciences (weather forecasting), and ML
(NLP) (Figure 2). For each task, we curate published benchmarks
and use www.green-algorithms.org to estimate the GHG emis-
sions (Experimental Section). For parameters independent of the
algorithm itself, we use average worldwide values, such as the
worldwide average power usage effectiveness (PUE) of 1.67[26]
and carbon intensity (CI) of 475 gCO2 e kWh−1 .[27]
2.1. Particle Physics Simulations
In particle physics, complex simulations are used to model the
passage of particles through matter. Geant4[28] is a popular toolkit
based on Monte-Carlo methods with wide-ranging applications,
such as the simulation of detectors in the Large Hadron Collider
and analysis of radiation burden on patients in clinical practice
or external beam therapy.[29–31] Meylan et al.[32] investigated the
biological effects of ionizing radiations on DNA on an entire hu-
man genome (6.4 × 109 nucleotide pairs) using GEANT4-DNA,
an extension of GEANT4.
To quantify the DNA damage of radiation, they run experi-
ments with photons of different energy, from 0.5 to 20 MeV. Each
© 2021 The Authors. Advanced Science published by Wiley-VCH GmbH
www.advancedsciencenews.com
Figure 1. The Green Algorithms calculator (www.green-algorithms.org).
experiment runs for three weeks to simulate 5000 particles (pro-
tons) using 24 processing threads and up to 10 GB of memory.
Using the Green Algorithms tool, and assuming an average CPU
power draw (such as the Xeon E5-2680, capable of running 24
threads on 12 cores), and worldwide average values for PUE and
CI, we estimated that a single experiment emits 49 465 gCO2 e.
When taking into account a PSF of 11, corresponding to the 11
different energy levels tested, the carbon footprint of such study
is 544 115 gCO2 e. Using estimates of car and air travel (Exper-
imental Section), 544 115 gCO2 e is approximately equivalent to
driving 3109 km (in a European car) or flying economy from New
Adv. Sci. 2021, 8, 2100707 2100707 (3 of 10)
www.advancedscience.com
York to San Francisco. In terms of carbon sequestration (Experi-
mental Section), it would take a mature tree 49 years to remove
the CO2 equivalent to the GHG emissions of this study from the
atmosphere (593 tree-months).
A common way to reduce the running time of algorithms is
to distribute the computations over multiple processing cores. If
the benefit in terms of time is well documented for each task,
as in[33] , the environmental impact is usually not taken into ac-
count. GEANT4 is a versatile toolbox; it contains an electromag-
netic package simulating particle transport in matter and high
energy physics detector response.[34] Schweitzer et al.[33] use a
© 2021 The Authors. Advanced Science published by Wiley-VCH GmbH
www.advancedsciencenews.com
Figure 2. Carbon footprint (gCO2 e) for a selection of algorithms, with and without their pragmatic scaling factor.
Figure 3. Effect of parallelization using multiple cores on run time and
carbon footprint using TestEm12 GEANT4 simulation.
standardized example, TestEm12,[35] to compare the perfor-
mances of different hardware configurations, from 1 to 60 cores
(i.e., a full Xeon Phi CPU). With the Green Algorithms tool, we
estimated the carbon footprint of each configuration (Figure 3),
which shows that increasing the number of cores up to 15 im-
proves both running time and GHG emissions. However, when
multiplying the number of cores further by 4 (from 15 to 60), the
running time is only halved, resulting in a twofold increase in
emissions, from 238 to 481 gCO2e. Generally, if the reduction in
running time is lower than the relative increase in the number of
Adv. Sci. 2021, 8, 2100707
www.advancedscience.com
cores, distributing the computations will worsen the carbon foot-
print. In particular, scientists should be mindful of marginal im-
provements in running time which have disproportionally large
effects on GHG emissions, as demonstrated by the gap between
30 and 60 cores in Figure 3. For any parallelized computation,
there is likely to be a specific optimal number of cores for mini-
mal GHG emissions.
2.2. Weather Forecasting
Weather forecasts are based on sophisticated models simulating
the dynamics between different components of the earth (such
as the atmosphere and oceans). Operational models face strin-
gent time requirements to provide live predictions to the pub-
lic, with a goal of running about 200–300 forecast days (FDs)
in one (wall clock) day.[36] Neumann et al.[36] present the perfor-
mances of two models in use for current weather forecasts: i) the
Integrated Forecast System (IFS)[37] used by the European Cen-
tre for Medium-Range Weather Forecasts (ECMWF) for 10-day
forecasts, and ii) the ICOsahedral Non-hydrostatic (ICON)[38] de-
signed by the German Weather Service (Deutscher Wetterdienst,
DWD) and whose predictions are used by more than 30 national
weather services.[39]
The configurations in daily use by the ECMWF include a su-
percomputer based in Reading, UK, which has a PUE of 1.45,[40]
while ICON is run on the German Meteorological Computa-
tion Centre (DMRZ)[41] based in Germany (PUE unknown). Neu-
mann et al.[36] ran their experiments on hardware similar to what
equips both facilities, “Broadwell” CPU nodes (Intel E5-2695v4,
36 cores) and minimum 64 GB memory per node. We utilize
these parameters for our CO2 e emission estimates. It is impor-
tant to note that ICON and IFS each solve slightly different prob-
lems, and therefore are not directly comparable.
2100707 (4 of 10) © 2021 The Authors. Advanced Science published by Wiley-VCH GmbH
www.advancedsciencenews.com
The DWD uses ICON with a horizontal resolution of 13 km[42]
and generates a FD in 8 min. Based on the experiments run by
Neumann et al.,[36] this requires 575 Broadwell nodes (20 700
CPU cores). We estimate that generating one FD emits 12 848
gCO2 e (14 tree-months). With a running time of 8 min per
FD, ICON can generate 180 FDs in 24 h. When taking into ac-
count this PSF of 180, we estimated that each day, the ICON
weather forecasting algorithm releases ≈2 312 653 gCO2 e, equiv-
alent to driving 13 215 km or flying from New York to San Fran-
cisco four times. In terms of carbon sequestration, the emissions
of each day of ICON weather forecast are equivalent to 2523 tree-
months.
At ECMWF, IFS makes 10-day operational weather forecasts
with a resolution of 9 km. To achieve a similar threshold of 180
FDs per day, 128 Broadwell nodes are necessary (4608 cores).[36,43]
Using the PUE of the UK ECMWF facility (1.45), we estimate the
impact of producing one FD with IFS to be 1660 gCO2 e. Using
a PSF of 180 for one day’s forecasts, we estimated emissions of
298 915 gCO2 e, equivalent to driving 1708 km or three return
flights between Paris and London. These emissions are equiva-
lent to 326 tree-months.
Furthermore, we modeled the planned scenario of the
ECMWF transferring its supercomputing to Bologna, Italy, in
2021.[44] Compared to the data center in Reading, the new data
center in Bologna is estimated to have a more efficient PUE of
1.27.[45] Prima facie this move appears to save substantial GHG
emissions; however, it is notable that the CI of Italy is 33% higher
than the UK.[46] Unless the sources of electricity for the data cen-
ter in Bologna are different from the rest of Italy and in the ab-
sence of further optimizations, we estimated that the move would
result in an 18% increase in GHG emissions from the ECMWF
(from 298 915 to 350 063 gCO2 e).
2.3. Natural Language Processing
In NLP, the complexity and financial costs of model training
are major issues.[17] This has motivated the development of lan-
guage representations that can be trained once to model the com-
plexity of natural language, and which could be used as input
for more specialized algorithms. The BERT (Bidirectional En-
coder Representations from Transformers)[47] algorithm is a field
leader which yields both high performance and flexibility: state-
of-the-art algorithms for more specific tasks are obtained by fine-
tuning a pre-trained BERT model, for example in scientific text
analysis[48] or biomedical text mining.[49] Yet, while the BERT
model is intended to avoid retraining, many data scientists, per-
haps understandably, continue to recreate or attempt to improve
upon BERT, leading to redundant and ultimately inefficient com-
putation as well as excess CO2 e emissions. Even with optimized
hardware (such as NVIDIA Volta GPUs), a BERT training run
may take three days or more.[50]
Using these optimized parameters, Strubell et al.[18] showed
that a run time of 79 h on 64 Tesla V100 GPUs was necessary to
train BERT, with a usage factor of the GPUs of 62.7%. With the
Greens Algorithms calculator, we estimated that a BERT training
run would emit 754 407 gCO2 e (driving 4311 km in a European
car; 1.3 flights from New York to San Francisco; and 823 tree-
months). When considering a conservative PSF of 100 for hyper-
Adv. Sci. 2021, 8, 2100707 2100707 (5 of 10)
www.advancedscience.com
parameters search, we obtain a carbon footprint of 75 440 740
gCO2 e.
While BERT is a particularly widely utilized NLP tool, Google
has also developed a chatbot algorithm, Meena, which was
trained for 30 days on a TPU-v3 Pod containing 2048 Tensor Pro-
cessing Unit (TPU) cores.[51] There is limited information on the
power draw of TPU cores and memory; however, the power sup-
ply of this pod has been estimated to be 288 kW.[52] Using a run
time of 30 days, assuming full usage of the TPUs and ignoring
memory power draw, the Greens Algorithms calculator estimated
that Meena training emitted 164 488 320 gCO2 e, which corre-
sponds to 179 442 tree-months or 71 flights between New-York
and Melbourne.
3. Discussion
The method and Green Algorithms tool presented here provides
users with a practical way to estimate the carbon footprint of their
computations. The method focuses on producing sensible esti-
mates with small overheads for scientists wishing to measure
the footprint of their work. Consequently, the online calculator
is simple to use and generalizable to nearly any computational
task. We applied the Green Algorithms calculator to a variety of
tasks, including particle physics simulations, weather forecast-
ing, and NLP, to estimate their relative and ongoing carbon emis-
sions. Real-world changes to computational infrastructures, such
as moving data centers, was also quantifiable in terms of carbon
footprint and was shown to be of substantive importance; for ex-
ample, moving data centers may attain a more efficient PUE but a
difference in CI may negate any efficiency gains, potentially mak-
ing such a move detrimental to the environment.
Our work substantially enhances and extends prior frame-
works for estimating the carbon footprint of computation. In
particular, we have integrated and formalized previously unclear
factors such as core usage and unitary power draw (per-core
or per-GB of memory). As a result, and as presented in the
Experimental Section, the carbon footprint of an algorithm can
be broken down to a small number of key, easily quantifiable
elements, such as number of cores, memory size and usage
factor. This reduces the burden on the user, who is not required
to either measure the power draw of hardware manually or use
a limited range of cloud providers for their computations. This
makes the method highly flexible in comparison to previous
work. Besides drawing attention to the growing issue of GHG
emissions of data centers, one of the benefits of presenting a
detailed open methodology and tool is to provide users with the
information they need to reduce their carbon footprint. Perhaps
the most important challenge in green computing is to make the
estimation and reporting of GHG emissions a standard practice.
This requires transparent and easy-to-use methodology, such
as the Green Algorithms calculator (www.green-algorithms.org)
and the open-source code and data presented here (see section
Code availability).
Our approach has a number of limitations. First, the carbon
footprint estimated is restricted to GHGs emitted to power com-
puters during a particular task. We do not perform a life cycle
assessment and therefore, do not consider the full environmen-
tal and social impact of manufacturing, maintaining, and dis-
posing of the hardware used, or the maintenance of the power
© 2021 The Authors. Advanced Science published by Wiley-VCH GmbH
www.advancedsciencenews.com
plants. Including these is impractical at scale and would greatly
reduce who can use the method. Besides, the conversion of the
impact of various GHG into CO2 e is commonly based on a 100-
year timescale; however, this is now debated as it can misrepre-
sent the impact of short-lived climate pollutants like methane[53]
and new standards may be needed in the future. Second, the TDP
may substantially underestimate power draw in some situations.
For example, when hyperthreading is used, the real power con-
sumption can be double the indicated TDP.[54] The TDP value
remains a sensible estimate of the base consumption of the pro-
cessor in most situations, but users using hyperthreading should
be aware of the impact on power consumption. Third, while the
power consumption from storage is usually minimal at the scale
of one computation, if central storage is constantly queried by
the algorithm (for example, to avoid overloading memory), this
can be an important factor in power draw; however, there are re-
sources which can be utilized if the algorithm is designed to be
heavily storage reliant.[55] Moreover, at the scale of the data cen-
ter, storage represents a significant part of electricity usage[55] and
research projects relying on large databases should separately
acknowledge the long-term carbon footprint of storage. Fourth,
while some averaging is necessary, the energy mix of a coun-
try varies by the hour. For example, the CI of South Australia,
which relies on wind and gas to produce electricity,[56] can vary
between 112 and 592 gCO2 e kWh−1 within one day, depending
on the quantity of coal-produced electricity imported from the
neighboring state of Victoria.[57] Although most regions are rel-
atively stable, these outliers may require a finer estimation. Our
online calculator uses averaged values sourced from government
reports.[46] Fifth, the PUE has some limitations as a measure of
data centers energy usage,[58,59] due to inconsistencies in ways
to calculate it. For example, reporting of PUE is highly variable
from yearly averages to best-case scenarios, such as in winter
when minimal cooling is required (as demonstrated by Google’s
quarterly results[60] ). Whether to include infrastructure compo-
nents such as security or on-site power generation is also source
of discrepancies between data centers.[61] Although some com-
panies present well-justified results, many PUEs have no or in-
sufficient justification. Furthermore, PUE is not defined when
computations are run on a laptop or desktop computer. As the
device is used for multiple tasks simultaneously, it is impossible
to estimate the power overhead due to the algorithm. In the cal-
culator, we use a PUE of 1 because of the lack of information,
but we caution this should not be interpreted as a sign of effi-
ciency. Even though discrepancies will remain, the widespread
adoption of an accurate, transparent, and certified estimation of
PUE, such as the ISO/IEC standard,[62] would be a substantial
step for the computing community. Sixth, the carbon emissions
in the section Results are based on manual curation of the liter-
ature. When parameters such as usage factor or PUE were not
specified, we made some assumptions (100% core usage, or us-
ing average PUE) that can explain differences between our esti-
mates and the real emissions. For best results, authors should
estimate and publish their emissions.
There are various, realistic actions one can take to reduce the
carbon footprint of their computation. Acting on the various pa-
rameters in Green Algorithms (see Experimental Section), is a
clear and easy way approach. Below, we describe a selection of
practical changes one can make:
Adv. Sci. 2021, 8, 2100707 2100707 (6 of 10)
www.advancedscience.com
3.1. Algorithm Optimization
Increasing the efficiency of an algorithm can have myriad ben-
efits, even apart from reducing its carbon footprint. Therefore,
we highly recommend this and foresee algorithm optimization
as one of the most productive, easily recognizable core activities
of green computing. While speed is an obvious efficiency gain,
part of algorithm optimization also includes memory minimiza-
tion. The power draw from memory mainly depends on the mem-
ory available, not the actual memory used,[63] and the memory
available is often the peak memory needed for one step of the al-
gorithm (typically a merge or aggregation). By optimizing these
steps, one can easily reduce energy consumption.
3.2. Reduce the Pragmatic Scaling Factor
Limiting the number of times an algorithm runs, especially those
that are power hungry, is perhaps the easiest way to reduce carbon
footprint. Relatedly, best practices to limit PSF (as well as finan-
cial cost) include limiting parameter fine-tuning to the minimum
necessary and building a small-scale example for debugging.
3.3. Choice of Data Center
Carbon footprint is directly proportional to data center efficiency
and the CI of the location. The latter is perhaps the parameter
which most affects total carbon footprint because of inter-country
variation, from under 20 gCO2 e kWh−1 in Norway and Switzer-
land to over 800 gCO2 e kWh−1 in Australia, South Africa, and
some states in the USA. To rigorously assess the impact of punc-
tually relocating computations, the marginal CI, rather than the
average one, should be used.[64] The marginal value depends on
which power plant would be solicited to meet the unexpected in-
creased demand. Although it would ideally be used, it varies by
the hour and is often not practical to estimate accurately at scale.
When the marginal CI is unknown, the average one (presented
in Experimental Section and Figure S2, Supporting information)
can be used by scientists as a practical lower bound estimate to
assess the benefit of moving computations. Indeed, due to the
low operating cost of renewable technologies, the marginal power
plants (which are the last one solicited) are generally high-carbon
technologies such as fuel or gas[64] which leads the marginal CI to
be higher than the average CI. Besides, if the move is permanent,
by relocating an HPC facility or using cloud computing for exam-
ple, then the energy needs are incorporated into utility planning
and the average CI is the appropriate metric to use. Data cen-
ter efficiency (PUE) varies widely between facilities but, in gen-
eral, large data centers optimize cooling and power supply, reduc-
ing the energy overhead and making them more efficient than
personal servers. Notably, a 2016 report estimated that if 80% of
small data centers in the USA were aggregated into hyperscale fa-
cilities, energy usage would reduce by 25%.[65] For users to make
informed choices, data centers should report their PUE and other
energy metrics. While large providers like Google or Microsoft
widely advertise their servers’ efficiency,[60,66] smaller structures
often do not. As highlighted here, cloud providers offer the oppor-
tunity to use efficient data centers in low-carbon countries, and
they can be greener alternatives to local data centers.
© 2021 The Authors. Advanced Science published by Wiley-VCH GmbH
www.advancedsciencenews.com
3.4. Offsetting GHG emissions
Carbon offsetting is a flexible way to compensate for carbon foot-
print. An institution or a user themself can directly support re-
ductions in CO2 or other GHGs, for example by sponsoring fuel-
efficient stoves in developing countries, reduction in deforesta-
tion or hydroelectric and wind-based power plants.[67,68] The pros
and cons of carbon offsetting are still debated due to the variety
of mechanisms and intricate international legislations and com-
peting standards. Therefore, we only present here an overview
and point interested scientists at some resources. Multiple inter-
national standards regulate the purchase of carbon credits and
ensure the efficiency of the projects supported.[69] Most of the
well-established standards are managed by non-profits and abide
by the mechanisms set in place by the Kyoto protocol (in partic-
ular Certified Emission Reduction)[70] and the PAS 2060 Carbon
Neutrality standard from the British Standards Institution.[71] Al-
though the primary aim is carbon offsetting, projects are often
also selected in line with the United Nations’ Agenda 30 for Sus-
tainable Development,[72] a broader action plan addressing in-
equalities, food security, and peace. Amongst the most popular
standards are the Gold Standard (founded by WWF and other
NGOs),[73] Verra (formerly Verified Carbon Standard),[74] and the
American Carbon Registry (a private voluntary GHG registry).[75]
In addition to direct engagement with these standards, platforms
like Carbon Footprint[67] select certified projects and facilitate the
purchase of credits.
4. Conclusions
The framework presented here is generalizable to nearly any
computation and may be used as a foundation for other aspects
of green computing. The carbon footprint of computation is sub-
stantial and may be affecting the climate. We therefore hope that
this new tool and metrics raise awareness of these issues as well
as facilitate pragmatic solutions which may help to mitigate the
environmental consequences of modern computation. Overall,
with the right tools and practices, we believe HPC and cloud com-
puting can be immensely positive forces for both improving the
human condition and saving the environment.
5. Experimental Section
The carbon footprint of an algorithm depends on two factors: the en-
ergy needed to run it and the pollutants emitted when producing such
energy. The former depends on the computing resources used (e.g., num-
ber of cores, running time, and data center efficiency) while the later, called
carbon intensity, depends on the location and production methods used
(e.g., nuclear, gas, or coal).
There are several competing definitions of “carbon footprint,” and in
this project, the extended definition from Wright et al.[76] was used. The cli-
mate impact of an event is presented in terms of carbon dioxide equivalent
(CO2 e) and summarizes the global warming effect of the GHG emitted in
the determined timeframe, here running a set of computations. The GHGs
considered were carbon dioxide (CO2 ), methane (CH4 ) and nitrous oxide
(N2 O);[77] these are the three most common GHGs of the “Kyoto bas-
ket” defined in the Kyoto Protocol[78] and represent 97.9% of global GHG
emissions.[79] The conversion into CO2 e was done using Global Warm-
ing Potential (GWP) factors from the Intergovernmental Panel on Climate
Change (IPCC)[77,80] based on a 100-year horizon (GWP100).
Adv. Sci. 2021, 8, 2100707 2100707 (7 of 10)
www.advancedscience.com
When estimating these parameters, accuracy and feasibility must be
balanced. This study focused on a methodology that could be easily and
broadly adopted by the community and therefore, restricts the scope of the
environmental impact considered to GHGs emitted to power computing
facilities for a specific task. Moreover, the framework presented requires
no extra computation, nor involves invasive monitoring tools.
Energy Consumption: An algorithm’s energy[81] needs were modeled
as a function of the running time, the number, type, and process time of
computing cores (CPU or GPU), the amount of memory mobilized, and
the power draw of these resources. The model further included the effi-
ciency of the data center,[82], which represents how much extra power is
necessary to run the facility (e.g., cooling and lighting).
Similar to previous works,[18,24] this estimate was based on the power
draw from processors and memory, as well as the efficiency of the data cen-
ter. However, the formula was refined and flexibility was added by including
a unitary power draw (per core and per GB of memory) and the processor’s
usage factor. The energy consumption E (in kWh) was expressed as:
E = t × (nc × Pc × uc + nm × Pm ) × PUE × 0.001 (1)
where t is the running time (hours), nc the number of cores, and nm the
size of memory available (gigabytes). uc is the core usage factor (between
0 and 1). Pc is the power draw of a computing core and Pm the power draw
of memory (Watt). PUE is the efficiency coefficient of the data centre.
The assumptions made regarding the different components are dis-
cussed below. It is previously shown that the power draw of a server moth-
erboard is negligible[83] and in a desktop computer, the motherboard han-
dles a multitude of tasks, which makes it impractical to assess the fraction
of power usage attributable to a specific algorithm. For these reasons, it
was decided not to include the motherboard’s power draw in this model.
Power Draw of the Computing Core: The metric commonly used to re-
port the power draw of a processor, either CPU or GPU, is its thermal
design power (TDP, in Watt) and is provided by the manufacturer. TDP
values frequently corresponded to CPU specifications which include mul-
tiple cores, thus here TDP values were normalized to per-core. While TDP
is not a direct measure of power consumption, rather the amount of heat a
cooling system dissipates during regular use—it is commonly considered
a reasonable approximation.
The energy used by the processor was the power draw multiplied by
processing time, scaled by the usage factor. However, processing time
could not be known a priori and, on some platforms, tracking could be
impractical at scale. Modeling exact processing time of past projects may
also necessitate re-running jobs, which would generate unnecessary emis-
sions. Therefore, when this processing time is unknown, the simplifying
assumption was made that core usage is 100% of run time (uc = 1 in
Equation (1)).
Power Draw from Memory: Memory power draw is mainly due to back-
ground consumption with a negligible contribution from the workload and
database size.[63] Moreover, the power draw is mainly affected by the total
memory allocated, not by the actual size of the database used, because the
load is shared between all memory slots which keeps every slot in a power-
hungry active state. Therefore, the primary factor influencing power draw
from memory is the quantity of memory mobilized, which simply requires
an estimation of the power draw per gigabyte. Measured experimentally,
this was estimated to be 0.3725 W per GB.[63,84]
For example, requesting 29 GB of memory draws 10.8 W, which is the
same as one core of a popular Core-i5 CPU. Figure S1, Supporting infor-
mation further compares the power draw of memory to a range of popular
CPUs.
Power Draw from Storage: The power draw of storage equipment
(HDD or SSD) varies significantly with workload.[55] However, in regu-
lar use, storage is typically solicited far less than memory and is mainly
used as a more permanent record of the data, independently of the task
at hand. The power draw of storage was estimated to be 0.001 W per GB
(Note S2, Supporting information). As above, by comparison, the power
draw of memory (0.3725 W per GB) and a Core-i5 CPU (10.8 W per core)
are more than two orders of magnitude greater. While the researcher over-
head for approximating storage usage may not be substantial, it is unlikely
© 2021 The Authors. Advanced Science published by Wiley-VCH GmbH
www.advancedsciencenews.com
to make a significant difference to overall power usage (and GHG emis-
sions) estimation. Therefore, the power consumption of storage was not
considered in this work.
Energy Efficiency: Data center energy consumption includes additional
factors, such as server cooling systems, power delivery components, and
lighting. The efficiency of a given data center can be measured by the Power
Usage Effectiveness,[85,61] defined as the ratio between the total power
drawn by the facility and the power used by computing equipment:
Ptotal
PUE = (2)
Pcompute
A data center PUE of 1.0 represents an ideal situation where all power
supplied to the building is utilized by computing equipment. The global
average of data centers was estimated as 1.67 in 2019.[26] While data cen-
ters with relatively inefficient PUE may not report it as such, some data
centers and companies have invested significant resources to bring their
PUEs as close to 1.0 as possible; for example, Google uses ML to reduce
its global yearly average PUE to 1.10.[60,86]
Carbon Intensity of Energy Production: For a given country and energy
mix, the carbon footprint in CO2 e represents the amount of CO2 with the
same global warming impact as the GHGs emitted, which simplifies the
comparison between different electricity production methods. The carbon
footprint of producing 1 kWh of energy (CI) varies significantly between
locations due to the broad range of production methods (Figure S2, Sup-
porting information), from 12 gCO2 e kWh−1 in Switzerland (mainly pow-
ered by hydro) to 880 gCO2 e kWh−1 in Australia (mainly powered by coal
and gas).[57,46] The 2020 CI values aggregated by Carbon Footprint[46] were
used; these production factors take into account the GHG emissions at the
power plants (power generation) as well as, when available, the footprint
of distributing energy to the data center.
Estimation of Carbon Footprint: The carbon footprint C (in gCO2 e) of
producing a quantity of energy E (in kWh) from sources with a CI (in gCO2 e
kWh−1 ) is then:
C = E × CI (3)
By putting together Equations (1) and (3), the long-form equation of
the carbon footprint Cis obtained:
C = t × (nc × Pc × uc + nm × Pm ) × PUE × CI × 0.001 (4)
CO2 e of Driving and Air Travel: gCO2 e was contextualized by estimat-
ing an equivalence in terms of distance travelled by car or by passenger
aircraft. Previous studies had estimated the emissions of the average pas-
senger car in Europe as 175 gCO2 e km−1[77,87] (251 gCO2 e km−1 in the
United States[88] ). The emissions of flying on a jet aircraft in economy class
were estimated between 139 and 244 gCO2 e km−1 per person, depending
on the length of the flight.[77] Three reference flights were used: Paris to
London (50 000 gCO2 e), New York to San Francisco (570 000 gCO2 e), and
New York to Melbourne (2 310 000 gCO2 e).[89]
CO2 Sequestration by Trees: Trees play a major role in carbon seques-
tration and although not all GHGs emitted could be sequestered, CO2
represents 74.4% of these emissions.[90] To provide a metric of reversion
for CO2 e, the number of trees needed to sequester the equivalent emis-
sions of a given computation was computed. The metric tree-months, the
number of months a mature tree needs to absorb a given quantity of CO2 ,
was defined. While the amount of CO2 sequestered by a tree per unit of
time depends on a number of factors, such as its species, size, or environ-
ment, it was estimated that a mature tree sequesters, on average, ≈11 kg
of CO2 per year,[91] giving the multiplier in tree-months a value close to 1
kg of CO2 per month (0.92 g).
Pragmatic Scaling Factor: Many analyses are presented as a single run
of a particular algorithm or software tool; however, computations are rarely
performed only once. Algorithms are run multiple times, sometimes hun-
dreds, systematically or manually, with different parameterizations. Sta-
tistical models may include any number of combinations of covariates,
Adv. Sci. 2021, 8, 2100707 2100707 (8 of 10)
www.advancedscience.com
fitting procedures, etc. It is important to include these repeats in the car-
bon footprint. To take into account the number of times a computation is
performed in practice, the PSF was defined, a scaling factor by which the
estimated GHG emissions are multiplied.
The value and causes of the PSF vary greatly between tasks. In
ML, tuning the hyper-parameters of a model requires hundreds, if not
thousands,[18] of runs, while other tools require less tuning and can some-
times be run a smaller number of times. As per published work or the
user’s own experience, the PSF should be estimated for any specific task;
besides, in Green Algorithms it is provided for, and recommended that,
each user estimate their own PSF.
Supporting Information
Supporting Information is available from the Wiley Online Library or from
the author.
Acknowledgements
L.L. and J.G. contributed equally to this work. L.L. was supported by
the University of Cambridge MRC DTP (MR/S502443/1). J.G. was sup-
ported by a La Trobe University Postgraduate Research Scholarship jointly
funded by the Baker Heart and Diabetes Institute and a La Trobe Uni-
versity Full-Fee Research Scholarship. This work was supported by core
funding from: the UK Medical Research Council (MR/L003120/1), the
British Heart Foundation (RG/13/13/30194; RG/18/13/33946) and the
NIHR Cambridge Biomedical Research Centre (BRC-1215-20014). This
work was also supported by Health Data Research UK, which is funded
by the UK Medical Research Council, Engineering and Physical Sciences
Research Council, Economic and Social Research Council, Department of
Health and Social Care (England), Chief Scientist Office of the Scottish
Government Health and Social Care Directorates, Health and Social Care
Research and Development Division (Welsh Government), Public Health
Agency (Northern Ireland), British Heart Foundation, and Wellcome. M.I.
was supported by the Munz Chair of Cardiovascular Prediction and Preven-
tion. This study was supported by the Victorian Government’s Operational
Infrastructure Support (OIS) program. The views expressed are those of
the authors and not necessarily those of the NHS, the NIHR or the De-
partment of Health and Social Care.
Conflict of Interest
The authors declare no conflict of interest.
Data Availability Statement
The data that support the findings of this study are openly avail-
able in a GitHub repository at https://github.com/GreenAlgorithms/
green-algorithms-tool.
Keywords
climate change, computational research, green computing
Received: February 20, 2021
Revised: March 22, 2021
Published online: May 2, 2021
[1] N. Jones, Nature 2018, 561, 163.
[2] A. Andrae, T. Edler, Challenges 2015, 6, 117.
© 2021 The Authors. Advanced Science published by Wiley-VCH GmbH
www.advancedsciencenews.com
[3] R. B. Alley, K. A. Emanuel, F. Zhang, Science 2019, 363, 342.
[4] The Event Horizon Telescope Collaboration, K. Akiyama, A. Alberdi,
W. Alef, K. Asada, R. Azulay, A.-K. Baczko, D. Ball, M. Baloković, J. Bar-
rett, D. Bintley, L. Blackburn, W. Boland, K. L. Bouman, G. C. Bower,
M. Bremer, C. D. Brinkerink, R. Brissenden, S. Britzen, A. E. Broderick,
D. Broguiere, T. Bronzwaer, D.-Y. Byun, J. E. Carlstrom, A. Chael, C.
Chan, S. Chatterjee, K. Chatterjee, M.-T. Chen, Y. Chen, I. Cho, et al.,
Astrophys. J. Lett. 2019, 875, L1.
[5] The Event Horizon Telescope Collaboration, K. Akiyama, A. Alberdi,
W. Alef, K. Asada, R. Azulay, A.-K. Baczko, D. Ball, M. Baloković, J. Bar-
rett, D. Bintley, L. Blackburn, W. Boland, K. L. Bouman, G. C. Bower,
M. Bremer, C. D. Brinkerink, R. Brissenden, S. Britzen, A. E. Broderick,
D. Broguiere, T. Bronzwaer, D.-Y. Byun, J. E. Carlstrom, A. Chael, C.
Chan, S. Chatterjee, K. Chatterjee, M.-T. Chen, Y. Chen, I. Cho, et al.,
Astrophys. J. Lett. 2019, 875, L3.
[6] The Event Horizon Telescope Collaboration, K. Akiyama, A. Alberdi,
W. Alef, K. Asada, R. Azulay, A.-K. Baczko, D. Ball, M. Baloković, J. Bar-
rett, D. Bintley, L. Blackburn, W. Boland, K. L. Bouman, G. C. Bower,
M. Bremer, C. D. Brinkerink, R. Brissenden, S. Britzen, A. E. Broderick,
D. Broguiere, T. Bronzwaer, D.-Y. Byun, J. E. Carlstrom, A. Chael, C.
Chan, S. Chatterjee, K. Chatterjee, M.-T. Chen, Y. Chen, I. Cho, et al.,
Astrophys. J. Lett. 2019, 875, L4.
[7] A. Buniello, J. A. L. MacArthur, M. Cerezo, L. W. Harris, J. Hayhurst,
C. Malangone, A. McMahon, J. Morales, E. Mountjoy, E. Sollis, D.
Suveges, O. Vrousgou, P. L. Whetzel, R. Amode, J. A. Guillen, H. S.
Riat, S. J. Trevanion, P. Hall, H. Junkins, P. Flicek, T. Burdett, L. A.
Hindorff, F. Cunningham, H. Parkinson, Nucleic Acids Res. 2019, 47,
D1005.
[8] D. Ben-Israel, W. B. Jacobs, S. Casha, S. Lang, W. H. A. Ryu, M.
de Lotbiniere-Bassett, D. W. Cadotte, Artif. Intell. Med. 2020, 103,
101785.
[9] D. M. Dimiduk, E. A. Holm, S. R. Niezgoda, Integr. Mater. Manuf.
Innov. 2018, 7, 157.
[10] G. Choy, O. Khalilzadeh, M. Michalski, S. Do, A. E. Samir, O. S. Pi-
anykh, J. R. Geis, P. V. Pandharipande, J. A. Brink, K. J. Dreyer, Radiol-
ogy 2018, 288, 318.
[11] S. Athey, in The Economics of Artificial Intelligence: An Agenda, Na-
tional Bureau of Economic Research, Inc., Cambridge 2018, p. 507.
[12] XSEDE Impact—Usage Statistics, https://portal.xsede.org/#/gallery
(accessed: March 2021).
[13] C. Reimsbach-Kounatze, OECD Digital Economy Papers, Vol. 155,
OECD Publishing, Paris 2009.
[14] S. Sarkar, S. Misra, IET Network 2016, 5, 23.
[15] K. Gai, M. Qiu, H. Zhao, L. Tao, Z. Zong, J. Network Comput. Appl.
2016, 59, 46.
[16] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT Press, Cam-
bridge, MA 2016.
[17] R. Schwartz, J. Dodge, N. A. Smith, O. Etzioni, Green AI. Commun.
ACM 2020, 63, 54.
[18] E. Strubell, A. Ganesh, A. McCallum, Energy and Policy Consider-
ations for Deep Learning in NLP, in Proceedings of the 57th An-
nual Meeting of the Association for Computational Linguistics, Flo-
rence, Italy, Jul. 2019, pp. 3645–3650, https://doi.org/10.18653/v1/
P19-1355.
[19] K. Jahnke, C. Fendt, M. Fouesneau, I. Georgiev, T. Herbst, M. Kaasi-
nen, D. Kossakowski, J. Rybizki, M. Schlecker, G. Seidel, T. Henning,
L. Kreidberg, H.-W. Rix, Nat. Astron. 2020, 4, 812.
[20] S. Portegies Zwart, Nat. Astron. 2020, 4, 819.
[21] A. R. H. Stevens, S. Bellstedt, P. J. Elahi, M. T. Murphy, Nat. Astron.
2020, 4, 843.
[22] C. Stoll, L. Klaaßen, U. Gallersdörfer, Joule 2019, 3, 1647.
[23] Cambridge Bitcoin Electricity Consumption Index (CBECI), https://
cbeci.org/contact/ (accessed: March 2021).
Adv. Sci. 2021, 8, 2100707 2100707 (9 of 10)
www.advancedscience.com
[24] A. Lacoste, A. Luccioni, V. Schmidt, T. Dandres, arXiv191009700 Cs
2019.
[25] P. Henderson, J. Hu, J. Romoff, E. Brunskill, D. Jurafsky, J. Pineau, J.
Mach. Learn. Res. 2020, 21, 1.
[26] A. Lawrence, Uptime Institute Blog 2019, https://journal.
uptimeinstitute.com/is-pue-actually-going-up/ (accessed: April
2020).
[27] Emissions—Global Energy & CO2 Status Report 2019—Analysis,
https://www.iea.org/reports/global-energy-co2-status-report-2019/
emissions (accessed: February 2020).
[28] S. Agostinelli, J. Allison, K. Amako, J. Apostolakis, H. Araujo, P. Arce,
M. Asai, D. Axen, S. Banerjee, G. Barrand, F. Behner, L. Bellagamba,
J. Boudreau, L. Broglia, A. Brunengo, H. Burkhardt, S. Chauvie, J.
Chuma, R. Chytracek, G. Cooperman, G. Cosmo, P. Degtyarenko, A.
Dell’Acqua, G. Depaola, D. Dietrich, R. Enami, A. Feliciello, C. Fergu-
son, H. Fesefeldt, G. Folger, et al., Nucl. Instrum. Methods Phys. Res.
2003, 506, 250.
[29] S. Jan, G. Santin, D. Strul, S. Staelens, K. Assié, D. Autret, S. Avner,
R. Barbier, M. Bardiès, P. M. Bloomfield, D. Brasse, V. Breton, P.
Bruyndonckx, I. Buvat, A. F. Chatziioannou, Y. Choi, Y. H. Chung,
C. Comtat, D. Donnarieix, L. Ferrer, S. J. Glick, C. J. Groiselle, D.
Guez, P.-F. Honore, S. Kerhoas-Cavata, A. S. Kirov, V. Kohli, M. Koole,
M. Krieguer, D. J. van der Laan, et al., Phys. Med. Biol. 2004, 49,
4543.
[30] S. Jan, D. Benoit, E. Becheva, T. Carlier, F. Cassol, P. Descourt, T. Fris-
son, L. Grevillot, L. Guigues, L. Maigne, C. Morel, Y. Perrot, N. Re-
hfeld, D. Sarrut, D. R. Schaart, S. Stute, U. Pietrzyk, D. Visvikis, N.
Zahra, I. Buvat, Phys. Med. Biol. 2011, 56, 881.
[31] D. Sarrut, M. Bardiès, N. Boussion, N. Freud, S. Jan, J.-M. Létang,
G. Loudos, L. Maigne, S. Marcatili, T. Mauxion, P. Papadimitroulas,
Y. Perrot, U. Pietrzyk, C. Robert, D. R. Schaart, D. Visvikis, I. Buvat,
Med. Phys. 2014, 41, 064301.
[32] S. Meylan, S. Incerti, M. Karamitros, N. Tang, M. Bueno, I. Clairand,
C. Villagrasa, Sci. Rep. 2017, 7, 11923.
[33] P. Schweitzer, S. Cipière, A. Dufaure, H. Payno, Y. Perrot, D. R. C. Hill,
L. Maigne, Sci. Program. 2015, 2015, 980752.
[34] J. Apostolakis, A. Bagulya, S. Elles, V. N. Ivanchenko, J. Jacquemier,
M. Maire, T. Toshito, L. Urban, J. Phys. Conf. Ser. 2010, 219, 032044.
[35] Geant4—an Object-Oriented Toolkit for Simulation in HEP, https://
gitlab.cern.ch/geant4/geant4/tree/edb408b5618b3b1cd3f40c5759a
a5da4aa56bb7b/examples/extended/electromagnetic/TestEm5
(accessed: June 2020).
[36] P. Neumann, P. Düben, P. Adamidis, P. Bauer, M. Brück, L. Kornblueh,
D. Klocke, B. Stevens, N. Wedi, J. Biercamp, Philos. Trans. R. Soc., A
2019, 377, 20180148.
[37] S. Keeley, Modelling and Prediction: Integrated Forecast Sys-
tem, https://www.ecmwf.int/en/research/modelling-and-prediction
(Published in 2013, accessed: April 2020).
[38] G. Zängl, D. Reinert, P. Rípodas, M. Baldauf, Q. J. R. Meteorol. Soc.
2015, 141, 563.
[39] Wetter und Klima—Deutscher Wetterdienst—Numerical weather
prediction models—ICON (Icosahedral Nonhydrostatic) Model,
https://www.dwd.de/EN/research/weatherforecasting/num_
modelling/01_num_weather_prediction_modells/icon_description.
html (accessed: June 2020).
[40] S. Baylis, 2011, https://doi.org/10.21957/CEMUB9F8, https://www.
ecmwf.int/node/17450 (accessed: April 2020).
[41] Wetter und Klima—Deutscher Wetterdienst—Information
technology—Data processing, DMRZ, https://www.dwd.de/EN/
aboutus/it/functions/Teasergroup/dataprocessing.html?nn=24864
(accessed: June 2020).
[42] The horizontal resolution represents the level of geographical detail
achieved when modelling the different weather phenomenon.
© 2021 The Authors. Advanced Science published by Wiley-VCH GmbH
www.advancedsciencenews.com
[43] J. A. P. León, Progress in using single precision in the IFS, https://
www.ecmwf.int/en/newsletter/157/meteorology/progress-using-
single-precision-ifs (Published in 2018, accessed: June 2020).
[44] G. Lentze, ECMWF signs contract with Atos for new supercom-
puter, https://www.ecmwf.int/en/about/media-centre/news/2020/
ecmwf -signs-contract-atos-new-supercomputer (accessed: April
2020).
[45] ECMWF, ECMWF Q&A, https://www.ecmwf.int/sites/default/files/
medialibrary/2019-02/Industry_day_-_QA.pdf (accessed: April
2020).
[46] International Electricity Factors, https://www.carbonfootprint.com/
international_electricity_factors.html (accessed: November 2020).
[47] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-training
of Deep Bidirectional Transformers for Language Understanding, in
Proceedings of the 2019 Conference of the North American Chapter
of the Association for Computational Linguistics: Human Language
Technologies, Volume 1 (Long and Short Papers), Minneapolis, Min-
nesota, Jun. 2019, pp. 4171–4186, https://doi.org/10.18653/v1/N19-
1423.
[48] I. Beltagy, K. Lo, A. Cohan, in Proc. 2019 Conf. Empir. Methods Nat.
Lang. Process. 9th Int. Jt. Conf. Nat. Lang. Process. EMNLP-IJCNLP,
Association For Computational Linguistics, Hong Kong, China 2019,
pp. 3615–3620.
[49] J. Lee, W. Yoon, S. Kim, D. Kim, S. Kim, C. H. So, J. Kang, Bioinformat-
ics 2020, 36, 1234.
[50] NVIDIA AI, BERT Meets GPUs—Future Vision—Medium,
https://web.archive.org/web/20190521104957/https://medium.
com/future-vision/bert-meets-gpus-403d3fbed848 (Published in
2019, accessed: April 2020).
[51] D. Adiwardana, M.-T. Luong, D. R. So, J. Hall, N. Fiedel, R. Thop-
pilan, Z. Yang, A. Kulshreshtha, G. Nemade, Y. Lu, Q. V. Le,
arXiv:2001.09977 2020.
[52] P. Teich, Tearing Apart Google’s TPU 3.0 AI Coprocessor,
https://www.nextplatform.com/2018/05/10/tearing-apart-googles-
tpu-3-0-ai-coprocessor/ (Published in 2018, accessed: April 2020).
[53] M. R. Allen, K. P. Shine, J. S. Fuglestvedt, R. J. Millar, M. Cain, D. J.
Frame, A. H. Macey, npj Clim. Atmos. Sci. 2018, 1, 16.
[54] I. Cutress, Why Intel Processors Draw More Power Than Ex-
pected: TDP and Turbo Explained, https://www.anandtech.com/
show/13544/why-intel-processors-draw-more-power-than-expected-
tdp-turbo (accessed: January 2020).
[55] E. Tomes, N. Altiparmak, in Proc. 2017 IEEE Int. Conf. Cluster Com-
puting, IEEE, Piscataway, NJ 2017, pp. 625–626.
[56] SA.GOV.AU—SA’s electricity supply and market, https:
//www.sa.gov.au/topics/energy-and-environment/energy-supply/
sas-electricity-supply-and-market (accessed: July 2020).
[57] B. Tranberg, O. Corradi, B. Lajoie, T. Gibon, I. Staffell, G. B. Andresen,
Energy Strategy Rev. 2019, 26, 100367.
[58] R. C. Zoie, R. D. Mihaela, S. Alexandru, in Proc. 2017 5th Int.
Symp. Electr. Electron. Eng. ISEEE 2017, IEEE, Piscataway, NJ 2017,
pp. 1–6.
[59] J. Yuventi, R. Mehdizadeh, Energy Build. 2013, 64, 90.
[60] Efficiency—Data Centers—Google, https://www.google.com/about/
datacenters/efficiency/ (accessed: January 2020).
[61] V. Avelar, D. Azevedo, A. French, E. N. Power, WhitePaper 2012,
https://datacenters.lbl.gov/sites/all/files/WP49-PUE%20A%
20Comprehensive%20Examination%20of%20the%20Metric_v6.pdf
(accessed: November 2020).
[62] I. S. O./I. E. C. 30134-2:2016(en), Information technology—Data
centres—Key performance indicators—Part 2: Power usage effec-
tiveness (PUE), https://www.iso.org/obp/ui/#iso:std:iso-iec:30134:
-2:ed-1:v1:en (accessed: November 2020).
[63] A. Karyakin, K. Salem, in Proc. 13th Int. Workshop Data Management
on New Hardware (DaMoN), ACM Press, Chicago 2017, pp. 1–9.
Adv. Sci. 2021, 8, 2100707 2100707 (10 of 10)
www.advancedscience.com
[64] R. McCarthy, C. Yang, J. Power Sources 2010, 195, 2099.
[65] A. Shehabi, S. Smith, D. Sartor, R. Brown, M. Herrlin, J. Koomey, E.
Masanet, N. Horner, I. Azevedo, W. Lintner, United States Data Center
Energy Usage Report, Lawrence Berkeley National Laboratory, Berke-
ley, USA 2016.
[66] Azure global infrastructure | Microsoft Azure, https://azure.
microsoft.com/en-gb/global-infrastructure/. (accessed: April 2020).
[67] Carbon Offset Projects, https://www.carbonfootprint.com/
carbonoffsetprojects.html (accessed: April 2020).
[68] Project Map—ClimateCare, https://climatecare.org/project-map/
(accessed: April 2020).
[69] A. Michaelowa, I. Shishlov, S. Hoch, P. Bofill, A. Espelage, Overview
and Comparison of Existing Carbon Crediting Schemes, Nordic Environ-
ment Finance Corporation (NICA), Helsinki, Finland 2019.
[70] Emissions Trading | UNFCCC, https://unfccc.int/process/the-kyoto-
protocol/mechanisms/emissions-trading (accessed: April 2020).
[71] PAS 2060 Carbon Neutrality, https://www.bsigroup.com/en-GB/
PAS-2060-Carbon-Neutrality/ (accessed: April 2020).
[72] Transforming Our World: The 2030 Agenda for Sustainable Develop-
ment, United Nations, New York 2015.
[73] The Gold Standard, https://www.goldstandard.org/ (accessed: April
2020).
[74] Verra, https://verra.org/ (accessed: April 2020).
[75] American Carbon Registry, https://americancarbonregistry.org/ (ac-
cessed: April 2020).
[76] L. A. Wright, S. Kemp, I. Williams, Carbon Manage. 2011, 2, 61.
[77] N. Hill, R. Bramwell, E. Karagianni, L. Jones, J. MacCarthy, S. Hinton,
C. Walker, B. Harris, Department for Business Energy & Industrial
Strategy (BEIS) 2020.
[78] Kyoto Protocol To The United Nations Framework Convention On Cli-
mate Change, United Nations, New York, USA 1998.
[79] H. Ritchie, M. Roser, Our World Data 2020, https://ourworldindata.
org/co2-and-other-greenhouse-gas-emissions (accessed: June
2020).
[80] Fourth Assessment Report—IPCC, https://www.ipcc.ch/
assessment-report/ar4/ (accessed: December 2020).
[81] Power (in Watt, W) measures the instantaneous draw of a compo-
nent. Energy (in kilowatt-hours, kWh) measures the power draw over
time and is obtained by multiplying the power draw by the running
time.
[82] By data centre, we mean the facility hosting the cores and memory,
which may not be a dedicated data centre.
[83] H. Geng, Data Center Handbook, John Wiley & Sons, Hoboken, USA
2014.
[84] C. Angelini, I. W. August 29 2014, Intel Core i7-5960X, -5930K And
-5820K CPU Review: Haswell-E Rises, https://www.tomshardware.
com/uk/reviews/intel-core-i7-5960x-haswell-e-cpu,3918-13.html (ac-
cessed: January 2020).
[85] C. L. Belady, C. G. Malone, Proc. ASME 2007 InterPACK Conf. collo-
cated with the ASME/JSME 2007 Thermal Engineering Heat Transfer
Summer Conf. , ASME, New York, USA 2010, pp. 751.
[86] J. Gao, Machine Learning Applications for Data Center Optimization,
2014.
[87] E. Helmers, J. Leitão, U. Tietge, T. Butler, Atmos. Environ. 2019, 198,
122.
[88] O. US EPA, Greenhouse Gas Emissions from a Typical Passen-
ger Vehicle, https://www.epa.gov/greenvehicles/greenhouse-gas-
emissions-typical-passenger-vehicle (accessed: January 2020).
[89] Carbon Footprint Calculator, https://calculator.carbonfootprint.com/
calculator.aspx?tab=3 (accessed: June 2020).
[90] H. Ritchie, M. Roser, Greenhouse gas emissions, https://
ourworldindata.org/greenhouse-gas-emissions (accessed: Novem-
ber 2020).
[91] H. Akbari, Environ. Pollut. 2002, 116, S119.
© 2021 The Authors. Advanced Science published by Wiley-VCH GmbH